Deep learning-based automated segmentation of neck abscesses on multiparametric MRI.
Authors
Affiliations (3)
Affiliations (3)
- Department of Radiology, University of Turku and Turku University Hospital, Kiinamyllynkatu 4-8, 20520 Turku, Finland.
- Turku PET Center, University of Turku and Turku University Hospital, Kiinamyllynkatu 4-8, 20520 Turku, Finland.
- Department of Radiology, University of Turku and Turku University Hospital, Kiinamyllynkatu 4-8, 20520 Turku, Finland; Medical Imaging Center, Department of Radiology, Tampere University and Tampere University Hospital, Tampere, Finland. Electronic address: [email protected].
Abstract
To develop and evaluate automated deep learning (DL) segmentation of acute neck abscesses on MRI and to assess agreement between automated and manual quantitative measurements relevant to severity assessment. In 226 patients with surgically confirmed neck abscesses from a single-center emergency MRI database, a DL segmentation model (nnU-Net v2) was trained using two input configurations: (1) post-contrast T1-weighted images (T1C) only and (2) a three-channel fusion of T1C, T2-weighted fat saturated (T2FS), and apparent diffusion coefficient (ADC) maps, with T2FS and ADC rigidly registered to T1C space. Performance was evaluated using five-fold cross-validation. Spatial overlap (Dice similarity coefficient [DSC]), volumetric agreement (intraclass correlation coefficient [ICC]), and quantitative feature agreement were compared between automated and manual segmentations. The fusion model achieved a mean DSC of 0.828 ± 0.124 across all 226 cases in five-fold cross-validation. The T1C-only model achieved a mean DSC of 0.813 ± 0.153 and was non-inferior to the fusion model (p = 0.005). Volumetric agreement was excellent for both models (fusion ICC = 0.970, T1C-only ICC = 0.933). Quantitative features showed good to excellent agreement. Automated DL segmentation of neck abscesses on multiparametric MRI is feasible and yields volumetric and quantitative measurements in good to excellent agreement with manual segmentation. The T1C-only model was non-inferior to the three-channel fusion model, offering a parsimonious single-sequence alternative that avoids registration-related confounds.